A NNALI DELLA
S CUOLA N ORMALE S UPERIORE DI P ISA Classe di Scienze
G IOVANNI A LESSANDRINI
E LENA B ERETTA
E DI R OSSET
S ERGIO V ESSELLA
Optimal stability for inverse elliptic boundary value problems with unknown boundaries
Annali della Scuola Normale Superiore di Pisa, Classe di Scienze 4
esérie, tome 29, n
o4 (2000), p. 755-806
<http://www.numdam.org/item?id=ASNSP_2000_4_29_4_755_0>
© Scuola Normale Superiore, Pisa, 2000, tous droits réservés.
L’accès aux archives de la revue « Annali della Scuola Normale Superiore di Pisa, Classe di Scienze » (http://www.sns.it/it/edizioni/riviste/annaliscienze/) implique l’accord avec les conditions générales d’utilisation (http://www.numdam.org/conditions). Toute utilisa- tion commerciale ou impression systématique est constitutive d’une infraction pénale.
Toute copie ou impression de ce fichier doit contenir la présente mention de copyright.
Article numérisé dans le cadre du programme Numérisation de documents anciens mathématiques
http://www.numdam.org/
Optimal Stability for Inverse Elliptic Boundary
Value Problems with Unknown Boundaries
GIOVANNI ALESSANDRINI - ELENA BERETTA - EDI ROSSET - SERGIO VESSELLA
Abstract. In this paper we study a class of inverse problems associated to elliptic boundary value problems. More
precisely,
those inverse problems in which the role of the unknown isplayed
by an inaccessible part of the boundary and the roleof the data is
played
by overdetermined boundary data for the elliptic equation assigned on theremaining,
accessible, part of the boundary. We treat the caseof
arbitrary
space dimension n > 2. Such problems arise inapplied
contextsof nondestructive testing of materials for either electric or thermal conductors, and are known to be ill-posed. In this paper we obtain essentially best possible
stability
estimates. Here, in the context ofill-posed
problems, stability means thecontinuous dependence of the unknown upon the data when additional a
priori
information on the unknown boundary (such as its regularity) is available.
Mathematics Subject Classification (2000): 35R30 (primary), 35R25, 35R35, 35B60, 31B20 (secondary).
1. - Introduction
In this paper we shall deal with two inverse
boundary
valueproblems.
Suppose Q
is a bounded domain in W withsufficiently
smooth bound- arya S2,
apart
ofwhich,
say I(perhaps
some interior connectedcomponent
of a S2 or some inaccessibleportion
of the exteriorcomponent
ofa03A9),
is notknown. This could be the case of an
electrically conducting specimen,
which ispossibly
defective due to the presence of interior cavities or of corrodedparts,
which are not accessible to directinspection.
See for instance[K-S-V].
The aimis to detect the presence of such defects
by
nondestructive methodscollecting
current and
voltage
measurements on the accessiblepart
A of theboundary
If we assume that the inaccessible
part
I of a 03A9 iselectrically insulated, then, given
a nontrivialfunction 1/1
onA, having
zero average(which represents
Work
supported
in part by MURST.Pervenuto alla Redazione il 21 settembre 1999 e in forma definitiva il 24
giugno
2000.Vol. XXIX (2000), pp. 755-806
the
assigned
currentdensity
on the accessiblepart
A of8Q),
we have that thevoltage potential
u inside Q satisfies thefollowing
Neumanntype boundary
value
problem
Here,
v is the exterior unit normal to aS2 and cr = denotes the knownsymmetric conductivity
tensor ant it is assumed tosatisfy
ahypothesis
of uniform
ellipticity.
Let us remark that the solution to( 1.1 a )-( 1.1 c)
isunique
up to an undetermined additive constant. In order to
specify
asingle solution,
we shall assume, from now on, the
following
normalization conditionSuppose,
now, that E is an open subset of which is contained inA,
andon which the
voltage potential
can be measured.Then,
the inverseproblem
consists of
determining
Iprovided
is known. This is the firstobject
of ourstudy
and we shall refer to it as the Inverse Neumann Problem(Neumann
case,for
short).
An allied
problem
is the one associated to the direct Dirichletproblem
Here,
asabove, I,
A are theinaccessible, respectively, accessible,
parts of and ar is theconductivity
tensorsatisfying
the samehypotheses.
Our secondobject
ofstudy
is the inverseproblem consisting
in the determination of I from theknowledge
of orVM’ where E C A is as above. We shall refer to it asthe Inverse Dirichlet Problem
(Dirichlet
case, forshort).
We believe that also thisproblem
may be of interest for concreteapplications
ofnondestructing testing,
for instance in thermal
imaging.
In this case, the inaccessibleboundary
Icould
represent
apriviledged
isothermalsurface,
such as a solidification front.Of course, it should be
kept
in mindthat, dealing
with thermal processes, theevolutionary
model based onparabolic,
rather thanelliptic, equations
is ingeneral
moreappropriate,
for related issues see, forinstance, [B-K-W], [Bi], [V 1 ] . However,
we trust that also apreliminary study
of astationary
modelmay be instructive.
Such two
problems,
the Neumann and Dirichlet cases, are known to beill-posed.
Indeed there areexamples
that showthat,
under apriori assumptions
on the unknown
boundary
Iregarding
itsregularity (up
to any finite order ofdifferentiability),
the continuousdependence (stability)
of I from the measured data in the Neumanncase, orVM’
VIE in the Dirichletcase) is,
atbest,
oflogarithmic type.
See[A12]
and also[Al-R].
The main purpose of this paper is to prove
stability
estimates ofloga-
rithmic
type (hence,
bestpossible)
for both the Neumann and Dirichlet cases,(Theorems 2.1, 2.2),
when the spacedimension n >
2 isarbitrary.
We recallthat,
for the case n =2,
resultscomparable
to ours have been found in[Be-V]
when o~ is
homogeneous
and in[R], [Al-R]
when o~ can beinhomogeneous
and also discontinuous. Other related results for the case of dimension two can
be found in
[Bu-C-Yl], [Bu-C-Y2], [Bu-C-Y3], [Bu-C-Y4], [An-B-J].
Let usalso recall
that, typically,
the above mentioned results are based onarguments related,
in various ways, tocomplex analytic methods,
which do not carry over thehigher
dimensional case.In the
sequel
of thisIntroduction,
we shall illustrate the new tools wefound necessary to
develop
andexploit
when n > 2. Butfirst,
let us commentbriefly
on the connection with another inverseproblem
which has becomequite popular
in the last ten years,namely
the inverseproblem
of cracks. On onehand there are
similarities,
in fact a crack can be viewed as acollapsed cavity,
that is a
portion
of surface inside theconductor,
such that ahomogeneous
Neumann condition like
(I,lc)
holds on the two sides of the surface. On the other hand there aredifferences,
for theuniqueness
in the crackproblem
atleast two
appropriate
distinct measurements are necessary[F-V],
whereas forour
problems,
either the Neumann or the Dirichlet case, anysingle
nontrivialmeasurement suffices for
uniqueness,
see for instance[Be-V]
for a discussionof the
uniqueness
issue. Let us also recall that for the crackproblem
indimensions
bigger
than two, various basicproblems regarding uniqueness
arestill unanswered.
See,
for the available results and for references[Al-DiB ] .
It istherefore clear that a
study
of thestability
for the crackproblem
in dimensionshigher
than two shallrequire
new ideas.Nonetheless,
the authors believe that thetechniques developed
heremight
be useful also in the treatment of the crackproblem.
The methods we use in this paper are based
essentially
on asingle unifying
theme:
Quantitative
Estimatesof Unique Continuation,
and we shallexploit
itunder various different
facets, namely
thefollowing
ones.(a) Stability
Estimatesof
Continuationfrom Cauchy
Data. Since we aregiven
the
Cauchy
data on E for a solution u to(l.la),
we shall need to evaluate how much apossible
error on suchCauchy
data can affect the interior values of u. Suchstability
estimates forCauchy problems
associated toelliptic equations
have been a centraltopic
ofill-posed problems
since thebeginning
of their modemtheory, [H], [Pul], [Pu2]. Here,
since one ofour
underlying
aims will be to treat ourproblems
underpossibly
minimalregularity assumptions,
we shall assume theconductivity
cr to beLipschitz
continuous
(this
is indeed the minimalregularity ensuring
theuniqueness
for the
Cauchy problem, [PI], [M]).
Ourpresent stability
estimates(Propo-
sitions
3.1, 3.2, 4.1, 4.2)
will elaborate oninequalities
due toTrytten [T]
who
developed
a method first introducedby Payne [Pal], [Pa2].
The ad-ditional
difficulty
encountered here will be that we shall need to compare solutions u 1, u 2 which are defined onpossibly
different domainsQ2
whose boundaries are known to agree on the accessible
part A only.
Letus recall that a similar
approach,
but restricted to thetopologically simpler
two-dimensional
setting,
hasalready
been used in[All], [Be-V].
We shallobtain
that,
if the error on the measurement on theCauchy
data issmall,
then for the Neumann case,
also IVUII I
issmall,
in anL2
average sense,on
S21 B Q2,
thepart
ofS21
which exceedsQ2.
And the same holds foron
SZ2 BQI (Propositions 3.1, 3.2).
In the Dirichlet case instead weshall prove that u 1 itself is small in
S21 B Q2,
and the same holds for u2 onS22 (Propositions 4.1, 4.2).
(b)
Estimatesof
Continuationfrom
the Interior. We shall also need interior average lower bounds on u and on itsgradient (Propositions 3.3, 4.3),
onsmall balls contained inside S2. Bounds of this
type
have been introduced in[Al-Ros-S,
Lemma2.2]
in the context of a different inverseboundary
value
problem.
The tools here involve another form ofquantitative unique continuation, namely
thefollowing.
(c)
ThreeSpheres Inequalities.
Also this one is a rather classical theme in connection withunique
continuation. Aside from the classical Hadamard’s three circlestheorem,
in the context ofelliptic equations
we recall Lan- dis[La]
andAgmon [Ag].
Under ourassumptions
ofLipschitz continuity
on a, our estimates
(see (5.47) below)
shall refer to differentialinequalities
on
integral
normsoriginally
due to Garofalo and Lin[G-L],
laterdeveloped by
Brummelhuis[Br]
and Kukavica[Ku].
(d) Doubling Inequalities
in the Interior. This rather recent tool has been intro- ducedby
Garofalo and Lin in the above mentioned paper[G-L].
Itprovides
an efficient method of
estimating
the local averagevanishing
rate of a so-lution to
(l.la).
Let us recall that it alsoprovides
a remarkablebridge
to the
powerful theory
ofMuckenhoupt weights [C-F]
and that this last connection has beencrucially
used in[Al-Ros-S]
and also in[V2].
The
last, fundamental,
appearance ofquantitative
estimates ofunique
con-tinuation is the
following.
(e) Doubling Inequalities
at theBoundary.
For our purposes it will be crucial to evaluate thevanishing
rate of Vu(in
the Neumanncase)
or of u(in
the Dirichlet
case)
near the inaccessibleboundary
I. Inparticular,
the factthat such a rate is not worse than
polynomial (Propositions 3.5, 4.5)
is anessential
ingredient
inproving
that thestability
for our inverseproblems
arenot worse than
logarithmic (see
theproof
of Theorem2.1).
Such evalua- tions onvanishing
rates nearI,
where anhomogeneous boundary
conditionapplies (either (I,lc)
or( 1.2. c) ),
can be obtainedby
the so called Dou-bling Inequalities
at theBoundary.
Thestudy
of suchinequalities
has beeninitiated
by Adolfsson,
Escauriaza andKenig [A-E-K]
and laterdeveloped by
Kukavica andNystrom [Ku-N]
and Adolfsson and Escauriaza[A-E].
Inparticular,
in[A-E]
suchinequalities
are proven, for the Neumannproblem,
when the
boundary
is 1smooth, and,
for the Dirichletproblem,
whenthe
boundary
issmooth,
where the modulus ofcontinuity w
is ofDini
type.
We shall useessentially
suchresults,
with theonly
differencethat, mainly
forsimplicity
ofexposition,
we shall assume, in the Dirichletcase, that
w (t)
=t",
0 a ~1,
that is of Holdertype (Propositions 3.5, 4.5).
Let us also recall that theconjecture
which has been left openby
the above mentioned papers is that the
Doubling Inequality
at the Bound-ary should hold true when the
boundary
isLipschitz. Hopefully,
if suchconjecture
were proven, then ourstability results,
Theorems2.1, 2.2, might
be
generalized
as follows. If I is apriori
known to beLipschitz
with asufficiently
smallLipschitz
constant, then thelogarithmic stability
estimatesof Theorems
2.1,
2.2 shouldapply
also to this case. If instead theLips-
chitz constant of I is
large,
then the bestpossible stability
estimatemight
be worse than
logarithmic.
This lastexpectation
is motivatedby
the factthat two
Lipschitz
surfaces withlarge Lipschitz
constant may bearbitrarily
close in the sense of the Hausdorff
distance,
butlocally they
need not tobe
representable
asgraphs
in a common referencesystem (see
Rondi[R]
for an
example).
If ithappens
that this is the case for the unknown bound- aries1,, 12,
then itmight
alsohappen
that estimates on the smallness ofI
inQI (in
the Neumann case, forinstance)
are worse thanloga-
rithmic. In fact from the
proofs
ofPropositions 3.2, 4.2, the importance
ofproving
thatII, 12
arelocally represented
asgraphs
in a common referencesystem
will become evident. Thisproperty
ofII, 12
will be referred toby saying
thatII, 12
are RelativeGraphs.
Sufficient conditionsguaranteeing
that the boundaries of the two domains
S22
are RelativeGraphs
willbe examined in
Proposition
3.6. As wealready mentioned,
in this paperwe intend to strive after
optimal
results underpossibly
minimal apriori assumptions
ofregularity (see i )
andiii)
in Section2). Moreover,
verygeneral assumptions
on the unknownboundary
I are made. It may have afinite,
butundetermined,
number of connected components, and no restric- tion isplaced
on theirtopology.
Furthermore we use asingle
measurementcorresponding
to oneboundary data, either 1/1
or g, that can beprescribed arbitrarily. Concerning
theirregularity,
theassumptions (2.7a), (2.8a)
arequite
loose andessentially correspond
to the natural ones in the treatment of the directproblems (1.1), (1.2) respectively.
Inaddition,
we shall re-quire
a bound on the oscillation character(frequency)
of1/1
or of g. This isexpressed
as a bound on a ratio of two norms: eitherin the Neumann case,
or
in the Dirichlet case .
Such control will be necessary in order to dominate the
vanishing
rates ofthe solutions in terms of
quantities
whichdepend only
on theprescribed
data. Notice that F may be
arbitrarily large,
but it isexpected
that theconstants in the estimates of Theorems
2.1,
2.2 may deteriorate as F - oo.The
plan
of the paper is as follows.In Section 2 we shall state the main Theorems
2.1, 2.2,
we also stateCorollary
2.3 whichprovides
a finerinterpretation
of thestability
estimates inthe
previous
theorems.Here,
instead ofestimating
the Hausdorff distance of the domainsS21, 522,
we shall estimate their distancelocally,
in terms of thegraph representation
of theirboundaries,
and alsoglobally, by viewing
them asimbedded differentiable manifolds with
boundary.
Sections 3 and 4 contain the
proofs
of Theorem 2.1 and Theorem2.2, respectively.
Theproofs
arepreceded by
the statements of variousauxiliary propositions (Propositions 3.1-3.6, Propositions 4.1-4.5).
Section 4 contains also theproof
ofCorollary
2.3.Section 5 contains the
proof
of thepropositions regarding
the estimates of continuation forCauchy problems,
andnamely Propositions 3.1, 3.2,
4.1 and 4.2.Such
proofs
areaccompanied by
some intermediate lemmas. Lemma 5.1 collectssome
regularity
results for the direct Neumannproblem.
Lemmas5.2,
5.3 dealwith the technical notion of
regularized
distance as introducedby
Lieberman.Section 6 contains the
proofs
ofPropositions 3.3,
4.3concerning
estimatesof continuation from the interior.
Section 7 contains all the
proofs concerning doubling inequalities. Namely,
the
proofs
ofPropositions 3.4, 4.4, dealing
with the interiordoubling inequalities,
the
proofs
ofPropositions 3.5, 4.5,
where the results of Adolfsson and Escauriazaare
adapted
to thepresent
purposes. Their result for the Dirichletproblem
issummarized in Lemma 7.1.
Section 8 deals with Relative
Graphs,
first in Lemma 8.1 we treat thegeneral
case ofLipschitz boundaries,
and we conclude with theproof
ofPropo-
sition 3.6.
2. - The main results
When
representing locally
aboundary
as agraph,
it will be convenient to use thefollowing
notation. For every x E IEBn we shall set x =(x’, xn),
wherex’ E E R.
DEFINITION 2.1. Let S2 be a bounded domain in Given a, 0
1,
we shall say that a
portion S
of 8Q is of classc1,a
with constants po, E >0, if,
for anyP E S,
there exists arigid
transformation of coordinates under whichwe have P = 0 and
where (p
is a function on CRI-1 satisfying
and
REMARK 2.1. To the purpose of
simplifying
theexpressions
in the various estimatesthroughout
the paper, we have found it convenient to scale all norms in such a way thatthey
aredimensionally equivalent
to their argument and coincide with the standard definition when the dimensionalparameter
poequals
1. Forinstance,
for any cp E we setwhere
Similarly,
we shall setand so on for
boundary
and trace norms suchas I i)
Apriori information
on the domain.Our main Theorems
2.1,
2.2 will be based on thefollowing assumptions
on the domain. Given po, M >
0,
we assume:Here,
and in thesequel, 10 1
denotes theLebesgue
measure of S2. We shalldistinguish
twononempty parts, A,
I in a S2 and we assumeHere,
interiors and boundaries are intended in the relativetopology
inMoreover we assume that we can select a
portion
within Asatisfying
forsome
PI
E ~and
also, denoting by
theportion
of of all x E such thatdist(x, I )
po,Regarding
theregularity
ofaS2, given
a,E,
0 a1,
E >0,
we assume that(2.5)
a S2 is of class with constants po, E .In
addition,
in Theorem2.1,
we shall also assume thefollowing (2.6)
1 is of class 1 with constants po, E .REMARK 2.2. Observe that
(2.5) automatically implies
a lower bound on thediameter of every connected
component
of a S2.Moreover, by combining (2.1 )
with
(2.5),
an upper bound on the diameter of S2 can also be obtained. Note also that(2.1), (2.5) implicitly comprise
an apriori
upper bound on the number of connectedcomponents
ofi i ) Assumptions
about theboundary
data.Let us set
(that
is:APO
= We shall assume thefollowing
on the Neumann data1/1 appearing
inproblem ( 1.1 )
and,
for agiven
constant F >0,
Concerning
the Dirichletdata g appearing
in(1.2),
we assumeAs noted
already
in Remark2.1,
norms aresuitably
scaled so to bedimensionally equivalent
to theirargument.
i i i ) Assumptions
about theconductivity.
The
conductivity a
is assumed to be agiven
function from Ilgn with valuesn x n
symmetric
matricessatisfying
thefollowing
conditions forgiven
constantsÀ,
A, 0 h1, A > 0,
for every
(ellipticity)
for every x, y E
(Lipschitz continuity)
In the
sequel,
we shall refer to the set of constantsE,
a,M, F, À,
A asto the a
priori
data.THEOREM 2.1. Let
521, Q2
be two domainssatisfying (2.1), (2.5).
LetAi, Ii,
i =
1, 2,
be thecorresponding
accessible and inaccessible partsof their
boundaries.Let us assume that
A
1 =A2
=A, 521, S22
lie on the same sideof
A and that(2.2)- (2.4)
aresatisfied by
bothpairs Ai,
Leth, 12 satisfy (2.6).
Let Ui E bethe solution to
( 1.1 )
when S2 =Qi,
i =1, 2,
and let(2.7), (2.9)
besatisfied. If, given
E >0,
we havethen we have
where co is an
increasing
continuousfunction
on[0, oo)
whichsatisfies
and
C,
1], C >0,
0 1] 1 are constantsonly depending
on the apriori
data.Here
d1-l
denotes the Hausdorff distance between bounded closed sets of II~n THEOREM 2.2.Let 521, Q2
andAi, Ii,
i =l, 2,
be as in Theorem 2.1. Let(2.1)- (2.5)
besatisfied.
Let Ui E be the solution to(1.2)
when S2 =Qi,
i =1, 2,
and let(2.8), (2.9)
besatisfied. If, given
E >0,
we havethen we have
where (J) is as in
(2.12)
and the constants C, r~, C >0,
0 17 1only depend
onthe a
priori
data.COROLLARY 2.3. Let the
hypotheses of
either Theorem 2.1 or Theorem 2.2 besatisfied.
There exist ro, 0 ro :5 po,only depending
on po, E, a,and Eo
>0, only depending
on the apriori data,
such thatif
E Eo thenfor
every P E aQ
U aQ2
there exists arigid transformation of
coordinates under which P = 0 andwhere (pl, are
c1,a functions
on C whichsatisfy, for
every~8,
0 P
a,where
when Theorem 2.1
applies,
when Theorem 2.2
applies,
(J) is as in
(2.12)
and K > 0only depends
on E, a and(J. Furthermore,
thereexists a
C 1’a diffeomorphism
F : Rn-* such thatF (Q2)
=S21 and for
everyfi, 0 p
ot,with K, lù as above. Here I d : denotes the
identity mapping.
3. - Proof of Theorem 2.1
Here and in the
sequel
we shall denoteby
G the connected component ofS21 nQ2
such that CG.
The
proof
of Theorem 2.1 is obtained from thefollowing
sequence ofpropositions.
PROPOSITION 3.1
(Stability
Estimate of Continuation fromCauchy Data).
Letthe
hypotheses of
Theorem 2.1, except(2.6),
besatisfied.
We havewhere (J) is an
increasing
continuousfunction
on[0, oo)
whichsatisfies
and C > 0
depends
on À, A, E, a and Monly.
DEFINITION 3.1. Let S2 be a bounded domain in W. We shall say that
a
portion S
of is ofLipschitz
class with constants po, E >0, if,
for anyP E S,
there exists arigid
transformation of coordinates under which we have P = 0 andwhere w
is aLipschitz
continuous function on Csatisfying
and
Here the
Co,
1 norm is scaledaccording
to theprinciples
stated in Remark2.1,
that isPROPOSITION 3.2
(Improved Stability
Estimate of Continuation fromCauchy Data).
Let thehypotheses of Proposition
3.1 holdand,
inaddition,
let us assume that there exist L > 0 and ro, 0 ro po, such that a G isof Lipschitz
class withconstants ro, L. Then
(3.1 )
holds with (J)given by
where y > 0 and C > 0
only depend
on À, A, E, a, M, L andpo/ro.
We shall denote
PROPOSITION 3.3
(Lipschitz Stability
Estimate of Continuation from the In-terior).
Let S2 be a domainsatisfying (2.1),
such that 8S2 isof Lipschitz
class withconstants po, E. Let U E
H1 (Q)
be the solution to(1.1), where * satisfies
and, for
agiven
constant F >0,
and Q
satisfies (2.9).
For every p > 0 and every Xo EQ4p,
we havewhere C > 0
depends
on À,A,
E, M, F andplpo only.
REMARK 3.1. Let us notice that
if 1Jr
satisfies(2.7a)-(2.7d),
then it also satisfies(3.4a)-(3.4c)
up topossibly replacing
F with amultiple cF,
wherec
only depends
on E. Infact,
forfunctions 1/1 satisfying (2.7c)
thefollowing equivalence
relations can be obtainedPROPOSITION 3.4
(Interior Doubling Inequality).
Let S2 be a domainsatisfy- ing (2.1 ),
such that a Q isof Lipschitz
class with constants po, E. Let u EH (Q)
bethe solution to
( 1.1 ), where 1/1 satisfies (3.4)
and asatisfies (2.9).
For every p > 0 and every Xo EQp,
we havefor
every r,fl
s. t. 1fJ
and 0 p , where C > 0 and K > 0depend
on À,A,
E, M, F andp/ po only.
PROPOSITION 3.5
(Doubling Inequality
at theBoundary).
Let Q be a domainsatisfying (2.1)
and(2.5).
Let us assume that the accessible and inaccessible partsA,
Iof
itsboundary satisfy (2.2)-(2.4)
and(2.6).
Let u E be the solutionto
(1.1)
and let(2.7)
and(2.9)
besatisfied.
Let Xo E I. For any r > 0 and any{3
> 1 we havewhere C > 0 and K > 0
depend
on À,A, E,
M and Fonly.
In the
sequel,
it will beexpedient
to introduce aquantity
which is aslight
variation of the Hausdorff distance between
SZ
I andS22 .
DEFINITION 3.2. We call
modified
distance betweenS21
I andS22
the numberNotice that we
obviously
havebut,
ingeneral, dm
does not dominate the Hausdorffdistance,
and indeed it does notsatisfy
the axioms of a distance function. This is made clearby
the
following example: Ql
=BI (0), S22
=B1 (o) B Bl/2(O).
In this casedm (S21, S22)
=0,
whereasd1t(f2I, S22)
=1 /2.
PROPOSITION 3.6
(Relative Graphs).
Let521, S22
be bounded domainssatisfy- ing (2.5).
There exist numbersdo,
ro,do
>0,
0 ro po,for
which the ratiosPO
I!1l. only depend
on a and E, such thatif we
have-
Po PO
then the
following facts
hold:i)
For every P E8 521,
up to arigid transformation of coordinates
which maps P into theorigin,
we havewhere ~p2 are on C
lRn-1 satisfying
for
everyP,
0fl
a , where C > 0only depends
on a,fl
and E.ii)
There exists an absolute constant C > 0 such thati i i ) Any
connected component Gof S21 I nQ2
hasboundary of Lipschitz
class withconstants ro, L, where ro is as above and L > 0
only depends
on a and E.PROOF OF THEOREM 2.1. Let us
denote,
forsimplicity, Let n
> 0 be such thatOur first
goal
is theproof
of thefollowing inequality
where C > 0 and K > 0
depend
onX, A, E,
M and Fonly.
As a pre-liminary step,
let us show that(3.15)
holds true when d isreplaced
withdm
=Q2),
thequantity
introduced in Definition 3.2. Let us assume, withno loss of
generality,
that there exists xo Eh
C such thatdist(xo, S22)
=dm.
From
(3.14)
weobviously
haveSuppose now dm
po.By Proposition 3.5, picking
r =dm, P
= we havewhere C > 0 and K > 0
depend
onX, A, E,
M and Fonly.
From(2.5)
we canfind a ball
Br (wo)
of radius r =compactly
contained inSZ
1 f12
Hence, applying Proposition
3.3 with p =r/4,
we havewhere C > 0
depends
on X,A, E,
M and Fonly.
From(3.16)-(3.18)
wederive
On the other
hand, when dm >
po,(3.19)
follows from(3.18)
and from the trivial estimatewith C
only depending
on E and M. Hence ~e haveproved
thatwhere C > 0 and K > 0
depend
onX, A, E,
M and Fonly.
With no loss of
generality,
let yo E be such thatdist(yo, Q2)
= d.Let us notice that in
general
yo needs not tobelong
toa 0 1,
see theexample
below Definition 3.2. For this reason it is necessary to
analyse
various differentcases
separately. Denoting by h
=dist(yo,
let usdistinguish
thefollowing
three cases:
where
do is
thenumber
introduced inProposition
3.6.If case
i )
occurs,taking
zo E 1 suchthat yo - zo - h,
we have thatd -
h ? 2,
so thatd 2dm
and(3.15)
follows from(3.21).
If case
i i )
occurs, let us setWe have that
By applying Proposition
3.4 with r =d i , B = d0 ,
i we havewhere C > 0 and K > 0
depend on X, A, E,
a, M and Fonly.
Since h >do
we can
apply Proposition
3.3 with p =~, obtaining
where C > 0
depends
onÀ, A, E,
a, M and Fonly.
From(3.24)
and(3.25)
we have
Let
If 17 fi,
thend
i~ ,
so that d =2d
iand
(3.15)
follows from(3.26). If, otherwise,
17 >~,
then(3.15)
followstrivially,
likewise we did in(3.20).
If case
i i i )
occurs, then ddo
andProposition
3.6applies,
so thatby (3.13)
and(3.19)
weagain
obtain(3.15).
Hence, by Proposition 3.1,
we obtainwhere C > 0
depends on X, A, E,
M and Fonly,
whereas K > 0depends
onthe same
quantities
and in addition on a. Thus we have obtained astability
estimate of
log-log type. Next, by (3.27),
we can find co >0, only depending
on X, A,
E,
a, M andF,
such that co then ddo. Therefore, by Proposition 3.6,
G satisfies thehypotheses
ofProposition
3.2. Hence in(3.15)
we may
replace
q with where co is as inPropo-
sition 3.2
(a
modulus ofcontinuity
oflog type)
and obtain(2.11), (2.12).
04. - Proof of Theorem 2.2 and of
Corollary
2.3Here and in the
sequel
we shall denoteby
G the connectedcomponent
ofS21 nQ2
suchthat EGG.
The
proof
of Theorem 2.2 is obtained from thefollowing
sequence ofpropositions,
whichclosely parallel Propositions
3.1-3.5.PROPOSITION 4.1
(Stability
Estimate of Continuation fromCauchy Data).
Letthe
hypotheses of
Theorem 2.2 besatisfied.
We havewhere (J) is an
increasing
continuousfunction
on[0, oo)
whichsatisfies
where C > 0
depends
on À,A,
E, a and Monly.
PROPOSITION 4.2
(Improved Stability
Estimate of Continuation fromCauchy Data).
Let thehypotheses of Proposition
4.1 holdand,
inaddition,
let us assumethat there exist L > 0 and ro, 0 ro po, such that a G is
of Lipschitz
class withconstants ro, L. Then
(4.1 )
holds with cogiven by
where y > 0 and C > 0
only depend
onk, A,E,
a, M, L andpo/ro.
PROPOSITION 4.3
(Lipschitz Stability
Estimate of Continuation from the In-terior).
Let S2 be a domainsatisfying (2.1),
such that a S2 isof Lipschitz
class withconstants po, E. Let U E
HI(Q)
be the solution to(1.2),
where gsatisfies
and, for
agiven
constant F >0,
and Q
satisfies (2.9).
For every p > 0 and every Xo EQ2p,
we havewhere C > 0
depends
onk, A, E, M, F andp/,oo only.
REMARK 4.1. Let us notice that
if g
satisfies(2.8a)-(2.8c),
then it also satisfies(4.4a)-(4.4b)
up topossibly replacing
F with amultiple cF,
where conly depends
on E. Infact,
forfunctions g satisfying (2.8b)
thefollowing equivalence
relations can be obtainedPROPOSITION 4.4
(Interior Doubling Inequality).
Let SZ be a domainsatisfy- ing (2.1 ),
such that a Q isof Lipschitz
class with constants po, E. Let u E be the solution to(1.2),
where gsatisfies (4.4)
and asatisfies (2.9).
For every p > 0 and every xo ES2 p,
we havefor
every r,f3
s.t. 1f3
and 0f3r
p , where C > 0 and K > 0depend
on À, A, E, M, F andp/,oo only.
PROPOSITION 4.5
(Doubling Inequality
at theBoundary).
Let S2 be a domainsatisfying (2.1 )
and(2.5).
Let us assume that the accessible andinaccessible parts A,
I
of
itsboundary satisfy (2.2)-(2.4).
Let U E be the solution to(1.2)
andlet
(2.8)
and(2.9)
besatisfied.
Let Xo E I. For any r > 0 andany P
> 1 we havewhere C > 0 and K > 0
depend
on À,A,
E, a, M and Fonly.
PROOF OF THEOREM 2.2.
By using
the trivial estimatethe
proof
is obtainedsimilarly
to theproof
of Theorem2.1,
up to obviouschanges.
DPROOF OF COROLLARY 2.3. We have that
(2.15)-(2.16)
followimmediately
from
(3.12), (2.12)
and either(2.11) (when
Theorem 2.1applies)
or(2.14) (when
Theorem 2.2applies).
Next,
let us prove(2.17).
We can find rl,h,
0 rl ro, 0 h ro,only depending
on a,E,
po, and a number Nonly depending
on a,E, M,
such that there exist
points Pl
E 1 andcylinders Cl,
I =1,..., N,
centeredat
Pl, having height
2h and basis a(n - I)-dimensional
disk of radius ri, such thatUN 1 CI
covers both8Qi
1 anda 522,
and eachCl
has axisalong
the direction labeledby xn
in the localrepresentation (2.15)
when P =Pl.
Moreover weassume
2Cl
C for every I. Here2Cl
denotes thecylinder
with double sizes and the same center. Noticethat, possibly replacing
Eoby
a smallernumber,
we may assume that thefunctions Vi
in(2.15) satisfy
for every
Let us fix I = 1 and let us define
FI :
R" as follows.Letting
x =(x’, xn )
suitable coordinates near P =
Pi,
where
Here: TJ,
0 ~ 1,
is a smooth function such that = 1 whenIx’l
rl,- 0
when 2r,,
andi (a , b; ~ ) :
R -+ R is anuniformly
smoothfunction for every a, b E
[-h/2, h/2] satisfying
for every S E
R,
for everyand also
t (a, b; b) = a .
Here c 1 is an absolute constant. For instance we can choose r as a suitable
smoothing
of thepiecewise
linear function whosegraph joins (-h, -h), (b, a),
(h, h),
within the square(-h, h)
x(-h, h)
and coincides with the bisector of the first and thirdquadrant
outside. Now we haveand hence
One can
verify
thatFI (Q2nCI)
=Fl (x)
= x for every x EaQI naQ2,
F =
FI
satisfies(2.17)
and also that ifS22
isreplaced
withFI(Q2),
then(2.16)
continues to hold. We may iterate this
procedure defining inductively analogous
maps
/~
which deform coordinates within thecylinder 2Cl
andreplacing
at eachstage
S22
withF’¡(Q2).
In the end we set F =FN
o... oFl.
F is an orientationpreserving c1,a diffeomorphism satisfying (2.17)
such that =a S21
and also F = I d outside of the fixed smallneighbourhood
ofa S21 given by U£ 12C’ .
Therefore
F ( S22 )
= 05. - Proofs of the estimates of continuation for
Cauchy problems
Throughout
thissection,
letQi, S22
be two domainssatisfying (2.1), (2.5).
Let
Ai, Ii, i
=1, 2,
be thecorresponding
accessible and inaccessible parts of their boundaries. Let us assume thatA
1 -A2
=A, 521, S22
lie on the sameside of A and that
(2.2)-(2.4)
are satisfiedby
bothpairs A~ , Ii.
We shall denote
It is clear that
for every , 4
LEMMA 5.1.
The following
Schauder type estimates holdwhere C > 0